Professional Documents
Culture Documents
Geostatistical Functional Data Analysis Wiley Series in Probability and Statistics 1St Edition Jorge Mateu Editor Full Chapter
Geostatistical Functional Data Analysis Wiley Series in Probability and Statistics 1St Edition Jorge Mateu Editor Full Chapter
Edited by
Jorge Mateu
University Jaume I of Castellon
Castellon, Spain
Ramón Giraldo
National University of Colombia
Bogota, Colombia
This edition first published 2022
© 2022 John Wiley & Sons Ltd
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted,
in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, except as
permitted by law. Advice on how to obtain permission to reuse material from this title is available at
http://www.wiley.com/go/permissions.
The right of Jorge Mateu and Ramón Giraldo to be identified as the authors of the editorial material in this
work has been asserted in accordance with law.
Registered Offices
John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, USA
John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex, PO19 8SQ, UK
Editorial Office
9600 Garsington Road, Oxford, OX4 2DQ, UK
For details of our global editorial offices, customer services, and more information about Wiley products visit
us at www.wiley.com.
Wiley also publishes its books in a variety of electronic formats and by print-on-demand. Some content that
appears in standard print versions of this book may not be available in other formats.
10 9 8 7 6 5 4 3 2 1
v
Contents
Index 424
xiii
List of Contributors
John Ensley
Gregory Bopp
Pennsylvania State University
Pennsylvania State University
Department of Statistics
Department of Statistics
USA
USA
Maria Franco-Villoria
Martha Bohorquez Università di Modena e Reggio Emilia
National University of Colombia Department of Economics
Department of Statistics “Marco Biagi”
Colombia Italy
and
CADS - Center for Analysis Decisions
and Society
Human Technopole
Italy
xvi
Foreword
Functional data analysis (FDA) is a branch of statistics that analyses data providing
information about curves, surfaces, or anything else varying over a continuum.
In its most general form, under an FDA framework each sample element is a func-
tion. The continuum over which these functions are defined is often time, but may
also be spatial location, wavelength, probability, etc. In the 20 years since the first
books and papers on this topic, this field of statistics has received the attention and
encouragement of researchers in statistics and many applied disciplines and has
become an important and dynamic area of modern statistics. Topics that have been
covered include descriptive techniques, statistical inference, multivariate and non-
parametric methods, regression, generalized linear models, time series, and spatial
statistics.
Modern technology has made it possible to obtain large spatial and spatiotem-
poral data sets, and poses the challenge of statistical modeling of such data. The
combination of spatial statistics with FDA has emerged as a key approach. This
book presents new theories and methods to define, describe, characterize, and
model functional data indexed in spatial or spatio-temporal domains. The main
focus is on functional data obtained under a geostatistical framework, where the
domain is fixed and continuous. Specific topics considered include kriging, clus-
tering, regression, and optimal sampling, moving on in the last part of the book to
spatiotemporal data. Some chapters also consider the treatment of functional data
on lattices.
When we wrote our original book on the subject in the 1990s, James Ramsay and
I hoped that we would encourage FDA as a way of thinking, not simply a collection
of techniques. It has therefore been very pleasing to see the development of the
field since then, and the abundance of research activity in the area has confirmed
our hopes. I would urge readers and researchers to raise their sights above any
specific methods, obviously important that they are, to ask how considering data
as functions changes and broadens our statistical horizons. Particularly in the new
era of data science, this concerns both what data can be collected and how they
can be analyzed. I am sure this book will make a valuable contribution in helping
them to do so.
Spatial statistics has developed rapidly during the last 30 years. We have seen
an interesting progress both in theoretical developments and in practical stud-
ies. Some early applications were in mining, forestry, and hydrology. It seems to
be honest to remark that the increasing availability of computer power and skill-
ful computer software has stimulated the ability to solve increasingly complex
problems. Clearly, these problems have some common elements: they were all
of a spatial nature. Some theory was available, for example the random function
theory as developed by Yaglom and others in the 1960s. But that was largely insuf-
ficient to find generic solutions for the whole class of problems, and hence, the
applications required a new theory. Thereupon some far-reaching theories have
been developed: image reconstruction, Markov random fields, point process statis-
tics, geostatistics, and random sets, to mention just a few. As a next stage, these
theories were applied successfully to new disciplinary problems leading to mod-
ifications and extensions of mathematical and statistical procedures. We there-
fore notice a general scientific process that has occurred in the field of spatial
statistics: well-defined problems with a common character were suddenly on the
agenda, and data availability and intensive discussion with practical and disci-
plinary researchers resulted in new theoretical developments. Often, it is difficult
to say which was first, and what followed, but we see different theoretical models
developed for different applications.
Spatial statistics has hence emerged as an important new field of science. One
of the peculiarities is its power for visualization. A common cold-water fear of
many statisticians and mathematicians to analyze images, to communicate their
Geostatistical Functional Data Analysis, First Edition. Edited by Jorge Mateu and Ramón Giraldo.
© 2022 John Wiley & Sons Ltd. Published 2022 by John Wiley & Sons Ltd.
2 1 Introduction to Geostatistical Functional Data Analysis
results by maps, and to have to trust information in pictures was overcome. It has
led to interesting theories and better and more objective procedures for dealing
with spatial variation. Following Wittgenstein, we could state that we needed some
geniuses to tackle the obvious. Now, many results of a spatial statistical analysis
could be communicated smoothly toward the nonstatistical audience, like a disci-
plinary scientist, a policy-maker, or an interested student. They, in turn, were able
to judge whether a problem was solved, whether a policy measure was relevant or
was inspired by the beautiful pictures expressing deep thoughts on relevant issues.
The role in policy-making may be once more stressed. It is known that many
policy-makers are inclined to make a decision on the basis of a well developed,
well organized, and well understandable figure. They find it (rightly so!) rather
boring to use long lists of statistical data. But as political decisions affect us all, it
puts another responsibility on the back of statisticians: to make statistically sound
maps. It is often hard to say what that should be, but at the very least, we should be
able to generate pictures, maps, and graphs that rely on good data and that show
important aspects for decision-making.
In this way, spatial statistics has become a refreshing wind in statistics. We do
not need to do well much longer on difficult equations, long lists of data, and tables
with simulated controlled scenarios. But, to be clear on the back of all these nice
pictures a sound science with sometimes difficult and tedious derivations and deep
thoughts are still required to make serious progress.
Spatial statistics recognizes and exploits the spatial locations of data when
designing for, collecting, managing, analyzing, and displaying such data. Spatial
data are typically dependent, for which there are classes of spatial models
available that allow process prediction and parameter estimation. Spatially
arranged measurements and spatial patterns occur in a surprisingly wide variety
of scientific disciplines. The origins of human life link studies of the evolution of
galaxies, the structure of biological cells, and settlement patterns in archaeology.
Ecologists study the interactions among plants and animals. Foresters and agri-
culturalists need to investigate plant competition and account for soil variations
in their experiments. The estimation of rainfall and of ore and petroleum reserves
is of prime economic importance. Rocks, metals, and tissue and blood cells
are all studied at a microscopic level. Geology, soil science, image processing,
epidemiology, crop science, ecology, forestry, astronomy, atmospheric science, or
simply any discipline that works with data collected from different spatial loca-
tions, need to develop models that indicate when there is dependence between
measurements at different locations. Spatiotemporal variability is a relatively new
area within Spatial Statistics, which explains the scarcity of space-time statistical
tools 20 years ago. There has been a growing realization in the last decade that
knowing where data were observed could help enormously in answering the
substantive questions that precipitated their collection. One of the most powerful
1.1 Spatial Statistics 3
tools for spatial data analysis is the map. For example, in military applications,
the battlespace is mapped for command and control. The sensors are both in situ
and remote, and they generate spatially distributed data of many different kinds.
Producing a statistically optimal map, together with measures of map uncertainty,
which is always up to date, is a complicated task. Once these types of statistical
problems are solved, a geographic information system, or GIS, is well suited to
forming the decision-making maps.
Spatial statistics can be considered a natural generalization of signal processing
to higher dimensions. In traditional signal processing, one has a signal dependent
on a scalar variable t, which may belong to a discrete set or which may be contin-
uous. Spatial statistics is concerned with cases in which t is a multidimensional
index of dimension d > 1. In most practical examples d = 2, though much of the
basic theory and methodology is the same whatever the dimension. Although the
models and methods of spatial statistics have not developed as rapidly as those for
one-dimensional signal processing, there have nevertheless been substantial new
developments in recent years. Standard and modern references on spatial statistics
include the books of [1–4] among others.
Following Cressie [5], spatial data can be thought of as resulting from observa-
tions on the stochastic process {Z(s) ∶ s ∈ D}, where D is possibly a random set in
ℝd . If we believe that the roots of statistical science are in data, we can classify spa-
tial areas according to the type of observations encountered. Thus, (i) if D is a fixed
subset of ℝd and Z(s) is a random vector at location s ∈ D, we are dealing with geo-
statistical data; (ii) if D is a fixed (regular or irregular) collection of countably many
points of ℝd and Z(s) is a random vector at location s ∈ D, we are dealing with lat-
tice data; (iii) if D is a point process in ℝd and Z(s) is a random vector at location
s ∈ D, we are dealing with point patterns; (iv) if D is a point process in ℝd and
Z(s) is itself a random set, we are dealing with spatial objects. Geostatistical-type
problems are distinguished most clearly from lattice-and point-pattern-type prob-
lems by the ability of the spatial index s to vary continuously over a subset of ℝd .
A space-time process can be denoted by {Z(s, t) ∶ s ∈ D(t), t ∈ T}, where each of
Z, D, and T is possibly random.
Spatial statistics is one of the major methodologies of environmental statistics.
Its applications include producing spatially smoothed or interpolated representa-
tions of air pollution fields, calculating regional average means or regional average
trends based on data at a finite number of monitoring stations, and performing
regression analyses with spatially correlated errors to assess the agreement
between observed data and the predictions of some numerical model. The notion
of proximity in space is implicitly or explicitly present in the environmental
sciences. Proximity is a relative notion, relative to the spatial scale of the scientific
investigation. When a spatial dimension is present in an environmental study, the
statistician’s job is to create a statistical framework within which one carries out
4 1 Introduction to Geostatistical Functional Data Analysis
that realistic covariance structures can be specified that accurately capture the
complicated dynamical processes as found in geophysical applications.
In the absence of a spatial component, there is a large class of time series models
that could be used to represent the temporal variability. These are dynamic in the
sense that they exploit the fact that time flows in only one direction, and so the
state of the process at the current time is related to what happened at previous
times. Thus, one might consider the space–time process as a collection of spa-
tially correlated time series in continuous space, or on a spatial lattice. Although
these approaches include dynamical structures, without a descriptive spatial com-
ponent one lacks the ability to perform spatial prediction at locations without
observations. If both temporal and spatial components are present, it is natural to
combine the temporally dynamic state-space approach and the spatially descrip-
tive approach. These models are referred to as space–time dynamic models.
Spatial interpolation is an essential feature of many GIS. It is a procedure
for estimating values of a variable at unsampled locations. A map with iso-
lines is usually the visual output of such a process and plays a crucial role in
decision-making. Based on Tobler’s law of geography, which stipulates that
observations close together in space are more likely to be similar than those
farther apart, the development of models attempting to represent the way close
observations are related can sometimes be very problematic. The approaches can
be divergent and may therefore lead to very different results. As a consequence,
an understanding of the initial assumptions and methods used is the key to the
spatial interpolation process.
Surprisingly, when spatial interpolation tools are integrated within GIS, they are
often implemented in such a way that users have no real choice in selecting the
best possible methods, and if they do have a choice, required input parameters
are sometimes fixed, without any possible way of modifying them. One reason
for the frequent blind use of spatial interpolation methods, and spatial statistics
in general, probably has its origins in teaching. Despite the large variety of its
applications, the discipline has been confined to those fields where it has seen its
major developments. The progress made in spatial statistics is therefore usually
presented only in journals dedicated to statistics, mining, and petroleum engi-
neering. As a consequence, GIS users who have a different technical background
often do not have an in-depth knowledge of such spatial interpolation techniques.
Furthermore, since the conventional tests used in basic statistics usually gener-
ate some kind of categorical answer, the prerequisite experience and statistical
knowledge necessary for the proper use of spatial interpolation techniques are
often discouraging to this type of users. Nevertheless, during the last few years,
the diversity of the applications of these methods has encouraged the publication
of new books and new case studies and has stimulated a number of conferences
on the subject.
1.2 Spatial Geostatistics 7
values at all the points in the domain D are considered, it can be seen as a reality
of an infinitely large set of r.v.s, one at each point in the domain, which is known
as spatial random function (synonyms: stochastic process, random field).
When s spans across the domain under study, D, we have a family of r.v.s,
{Z(s), s ∈ D}, which constitutes a spatial random field (r.f.).
This methodological decision is one of the cornerstones of geostatistics: the
regionalized variable is interpreted as a realization of a spatial r.f. At this point,
we must state that the regionalized variable is often highly locally irregular
(which makes it impossible to represent using a deterministic mathematical
function) and has a certain spatial organization or structure. The probabilistic
approach, or probabilistic geostatistics, which interprets the regionalized variable
as a realization of a r.f., can take into account all the aspects of regionalization
mentioned above, because, as stated in page 55 of [7]:
i) At each location s, Z(s) is a r.v. (hence, the erratic aspect).
ii) For any given set of points s1 , s2 , … , sk , the r.v.s Z(s1 ), Z(s2 ), … , Z(sk ) are linked
by a network of spatial correlations responsible for the similarity of the values
they take (hence the structured aspect).
Let Z(s) be a r.f. and let us consider the set of points (s1 , … , sk ). Then, the
r.f. Z(s) is characterized by its k-dimensional distribution function. The set of
k-dimensional distribution functions for all values of k and all possible choices of
(s1 , … , sk ) in the domain is called the spatial law of probability.
For a given r.f., Z(s), the k-dimensional distribution function F(z(s1 )), … , (z(sk ))∶
ℝd → [0, 1] is defined as
FZ1 ,…,Zk (z(s1 )), … , (z(sk )) = P[Z(s1 ) ≤ z(s1 ), … , Z(sk ) ≤ z(sk )]. (1.1)
In linear geostatistics, it is enough to know the first two moments of the distribu-
tion of Z(s). What is more, in most practical applications, the available information
does not allow to infer higher-order moments.
The expectation, expected value or first-order moment of a r.f. is defined as a
nonrandom function of s that coincides at each point with the expectation of the
r.v. at that point 𝜇(s) = E(Z(s)), where 𝜇(si ) = E(Z(si )), ∀i ∈ ℕ. It is also called the
drift of the r.f., especially when it varies with location.
The variance of a r.f. is defined as a nonrandom function of s that coincides at
each point with the variance of the r.v. at that point, i.e. V(s) = V(Z(s)), where
V(si ) = V(Z(si )), ∀i ∈ ℕ.
The covariance function of a r.f. is defined as a nonrandom function of si and sj ,
such that for any pair of values (si , sj ) coincides with the covariance between the
r.v. at those two points
( )
C(si , sj ) = C(Z(si ), Z(sj )) = E (Z(si ) − 𝜇(si ))(Z(sj ) − 𝜇(sj )) , ∀si , sj ∈ D.
(1.2)
1.2 Spatial Geostatistics 9
The variogram of the r.f. is defined as the variance of the first differences of the r.f.
( )
2𝛾(si − sj ) = V (si ) − Z(sj ) , ∀si , sj ∈ D. (1.3)
● The expectation exists and is constant, and therefore does not depend on the
location s
● The covariance exists for every pair of r.v.s, Z(s) and Z(s + h), and only depends
on the vector h that joins the locations s and (s + h)
In light of Eqs. (1.4) and (1.6), the second-order stationarity hypotheses can
be interpreted as if the regionalized variable takes values that fluctuate around
a constant value (the mean), and the variation of these fluctuations is the same
everywhere in the domain.
In some cases, in order to model the spatial dependence of second-order station-
ary r.f.s, the correlogram, or correlation function, is used instead of the covariogram,
and is defined as
C(h)
Corr(Z(s), Z(s + h)) = = 𝜌(h). (1.7)
C(𝟎)
In case of second-order stationarity, the covariance function and the semi-
variogram are equivalent when it comes to defining the structure of spatial
1.2 Spatial Geostatistics 11
C((si , ti ), (sj , tj )) only depends on the distance between the locations (si and sj ) and
the times ti and tj . And the spatiotemporal r.f. Z(s, t) is said to have a temporarily
stationary covariance function if, for any two pairs (si , ti ) and (sj , tj ) on ℝd × ℝ,
the covariance C((si , ti ), (sj , tj )) only depends on the distance between the times
(ti and tj ) and the spatial locations si and sj . If the spatiotemporal r.f. Z(s, t) has a
stationary covariance function in both spatial and temporal terms, then it is said
to have a stationary covariance function. In this case, the covariance function can
be expressed as
C((si , ti ), (sj , tj )) = C(h, u) (1.11)
with h = si − sj and u = ti − tj the distances in space and time, respectively.
A spatiotemporal r.f. Z(s, t) has a separable covariance function if there is a purely
spatial covariance function Cs (si , sj ) and a purely temporal covariance function
Ct (ti , tj ) such that
C((si , ti ), (sj , tj )) = Cs (si , sj )Ct (ti , tj ) (1.12)
for any pair of spatiotemporal locations (si , ti ) and (sj , tj ) ∈ ℝd × ℝ.
A spatiotemporal r.f. Z(s, t) has a fully symmetrical covariance function if
C((si , ti ), (sj , tj )) = Cs (si , tj )Ct (sj , ti ) (1.13)
for any pair of spatiotemporal locations (si , ti ) and (sj , tj ) ∈ ℝ × ℝ. d
constructed from the random variables Z(si , ti ). As in the spatial case, spatiotempo-
ral kriging equations will depend on the degree of stationarity attributed to the r.f.
that supposedly generates the observed realization. The most widely used kriging
techniques in the spatiotemporal case are simple spatiotemporal kriging, ordi-
nary spatiotemporal kriging, and universal spatiotemporal kriging. In the case of
simple spatiotemporal kriging, we assume that Z(s, t) is a second-order stationary
spatiotemporal r.f., with a constant and known mean 𝜇(s, t), constant and known
variance C(𝟎, 0), and a known covariance function C(h, u). The kriging equations
(n equations with n unknown elements) are of the form
∑
n
( ) ( )
𝜆j C si − sj , ti − tj = C si − s0 , ti − t0 , ∀i = 1, … , n (1.28)
j=1
1.3 Spatiotemporal Geostatistics 17
from which we obtain the values 𝜆i that minimize the prediction error variance,
which is given by
[ ] ∑
n
V Z ∗ (s0 , t0 ) − Z(s0 , t0 ) = C(𝟎, 0) − 𝜆i C(si − s0 , ti − t0 ). (1.29)
i=1
In the case of ordinary spatiotemporal kriging, the constant mean 𝜇(s, t) is not
known, and the covariance function C(h, u) is known, under second-order sta-
tionarity. In the case of an intrinsic r.f., the variance is unbounded. In these two
cases, simple kriging cannot be performed as the mean cannot be subtracted. We
must therefore impose a condition of unbiasedness. In these situations, ordinary
spatiotemporal kriging equations can be expressed, in the first case, in terms of
the covariance function, and in the second case, in terms of the semivariogram, as
there is no covariance at the origin.
In the universal kriging approach, assume Z(s, t) is a r.f. with drift, and so the
mean of the r.f. is not constant, but depends on the pairs (s, t). In this situation,
the so-called “condition of unbiasedness” is affected substantially. In this case, the
r.f. can be disaggregated into two components: one deterministic 𝜇(s, t) and the
other stochastic e(s, t) which can be treated as an intrinsically stationary r.f. with
zero expectation, E [e(s, t)] = 0
Z(s, t) = 𝜇(s, t) + e(s, t). (1.30)
We can assume that the mean, even unknown, can be expressed locally by
∑
p
𝜇(s, t) = ah fh (s, t), (1.31)
h=1
{ }
where fh (s, t), h = 1, … , p are p known functions, ah constant coefficients, and
p the number of terms used in the approximation. It must be taken into account
that this expression is only valid locally. In this case, the equations that yield the
prediction of the weights are obtained from the prediction error conditions of zero
expectation and minimum variance.
[33, 43], criminology [44], economy [45], medicine [46], meteorology [47],
oceanography [48], psychology [49], and veterinary [50], among others, are
areas where this relatively recent and novel field of statistics is useful. Ullah
and Finch [51] presented a nice compilation of case studies analyzed from a
FDA perspective. To apply FDA to a real dataset, there is a need for appropriate
software with up-to-date methodological implementation and easy addition of
new theoretical developments [52]. R R Core Team [53] and python libraries [52]
can be used for this purpose.
In the geostatistical setting, kriging and cokriging methods have been also
extended to deal with functional data, and methods such as ordinary kriging for
functional data (OKFD), continuous time-varying kriging for functional data
(CTVKFD), and functional kriging total model (FKTM) are now easily available.
The simplest predictor abovementioned is OKFD, where each curve is weighted
by a scalar parameter. The second option (CTVKFD) is founded in the theory
of the functional linear concurrent (pointwise) regression model. In this case,
the parameters are also curves. Finally, the predictor (FKTM) considers double
indexed parameters. Here, in order to carry out the prediction at a particular
time, each observed curve is weighted by a functional parameter. This approach
follows the philosophy of the functional linear total model. Recently, in a similar
framework, these methods have been adapted for Hilbert Spaces.
All the predictors before referenced are based on the assumption that the mean
function is homogeneous into the region of interest. However, in practice, we often
found realizations of nonstationary functional processes (because there exists an
explicit spatial trend). To give solution to the problem of spatial prediction of func-
tional data in the absence of stationarity some alternatives have been proposed. All
of these have arisen as extensions to the functional framework of some classical
kriging methods for nonstationary data.
Geostatistical FDA has been a field of constant growing in the last years. New
modeling requirements in this area are opening many new research avenues. The
definition of predictors considering realizations of multivariate (possibly nonsta-
tionary) functional random fields or the spatial prediction of data belonging to
Riemannian manifolds are only a few new developments which indicate that there
is still a long way ahead in this area of statistics. This book shows a state-of-the-art
with recent contributions in this new environment of the spatial and functional
modeling.
New theories emerge every day that allow adapting and extending traditional
statistical methods to the treatment of functional data. However, there is still a
long way ahead, given the wide range of theoretical and applied possibilities not yet
explored. There are undoubtedly many challenges for the statistical community in
this area.
22 1 Introduction to Geostatistical Functional Data Analysis
References
38 Wang, J., Chiou, J., and Müller, H. (2016). Functional data analysis. Annual
Review of Statistics and Its Application 3: 257–295.
39 Zhao, X., Marron, J., and Wells, M. (2004). The functional data analysis view
of longitudinal data. Statistica Sinica 14 789–808.
40 Delicado, P., Giraldo, R., Comas, C., and Mateu, J. (2010). Statistics for spatial
functional data: some recent contributions. Environmetrics: The Official Journal
of the International Environmetrics Society 21(3–4): 224–239.
41 Giraldo, R., Caballero, W., and Camacho-Tamayo, J. (2018). Mantel test for spa-
tial functional data. Advances in Statistical Analysis 102 (1): 21–39.
42 Cremona, M., Xu, H., Makova, K. et al. (2019). Functional data analysis for
computational biology. Bioinformatics 35 (17): 3211–3213.
43 Lopez-Pintado, S. and Qian, K. (2021). A depth-based global envelope test
for comparing two groups of functions with applications to biomedical data.
Statistics in Medicine. https://doi.org/10.1002/sim.8861.
44 Park, K., Suk, H., Hwang, H., and Lee, J.H. (2013). A functional analysis of
deception detection of a mock crime using infrared thermal imaging and the
concealed information test. Frontiers in Human Neuroscience 7: 70.
45 Frois Caldeira, J., Gupta, R., Suleman, M., and Torrent, H. (2020). Forecasting
the term structure of interest rates of the BRICS: evidence from a nonparamet-
ric functional data analysis. In: Emerging Markets Finance and Trade, 1–18.
https://doi.org/10.1080/1540496X.2020.1808458.
46 Strzalkowska-Kominiak, E. and Romo, J. (2021). Censored functional data for
incomplete follow-up studies. Statistics in Medicine. https://doi.org/10.1002/sim
.8930/
47 Suhaila, J., Jemain, A., Hamdan, M., and Zin, W. (2011). Comparing rainfall
patterns between regions in Peninsular Malaysia via a functional data analysis
technique. Journal of Hydrology 411 (3–4): 197–206.
48 Assunção, R., Silva, A., Roy, A. et al. (2020). 3D characterisation of the
thermohaline structure in the southwestern tropical Atlantic derived from
functional data analysis of in situ profiles. Progress in Oceanography 187:
102399.
49 Levitin, D., Nuzzo, R., Vines, B., and Ramsay, J. (2007). Introduction to func-
tional data analysis. Canadian Psychology 48 (3): 135.
50 Gruen, M., Alfaro-Córdoba, M., Thomson, A. et al. (2017). The use of func-
tional data analysis to evaluate activity in a spontaneous model of degenerative
joint disease associated pain in cats. PLoS ONE 12 (1): e0169576.
51 Ullah, S. and Finch, C. (2013). Applications of functional data analysis: a
systematic review. BMC Medical Research Methodology 13 (1): 1–12.
References 25
Part I